Chinese Journal of Lasers, Год журнала: 2024, Номер 51(18), С. 1800001 - 1800001
Опубликована: Янв. 1, 2024
Chinese Journal of Lasers, Год журнала: 2024, Номер 51(18), С. 1800001 - 1800001
Опубликована: Янв. 1, 2024
Light Science & Applications, Год журнала: 2024, Номер 13(1)
Опубликована: Сен. 20, 2024
Язык: Английский
Процитировано
24Nature Communications, Год журнала: 2025, Номер 16(1)
Опубликована: Янв. 29, 2025
Язык: Английский
Процитировано
11Chemical Reviews, Год журнала: 2025, Номер unknown
Опубликована: Янв. 2, 2025
Recent breakthroughs in brain-inspired computing promise to address a wide range of problems from security healthcare. However, the current strategy implementing artificial intelligence algorithms using conventional silicon hardware is leading unsustainable energy consumption. Neuromorphic based on electronic devices mimicking biological systems emerging as low-energy alternative, although further progress requires materials that can mimic function while maintaining scalability and speed. As result their diverse unique properties, atomically thin two-dimensional (2D) are promising building blocks for next-generation electronics including nonvolatile memory, in-memory neuromorphic computing, flexible edge-computing systems. Furthermore, 2D achieve biorealistic synaptic neuronal responses extend beyond logic memory Here, we provide comprehensive review growth, fabrication, integration van der Waals heterojunctions optoelectronic devices, circuits, For each case, relationship between physical properties device emphasized followed by critical comparison technologies different applications. We conclude with forward-looking perspective key remaining challenges opportunities applications leverage fundamental heterojunctions.
Язык: Английский
Процитировано
5Science, Год журнала: 2023, Номер 382(6676), С. 1297 - 1303
Опубликована: Ноя. 23, 2023
Recent successes in deep learning for vision and natural language processing are attributed to larger models but come with energy consumption scalability issues. Current training of digital deep-learning primarily relies on backpropagation that is unsuitable physical implementation. In this work, we propose a simple neural network architecture augmented by local (PhyLL) algorithm, which enables supervised unsupervised networks without detailed knowledge the nonlinear layer's properties. We trained diverse wave-based vowel image classification experiments, showcasing universality our approach. Our method shows advantages over other hardware-aware schemes improving speed, enhancing robustness, reducing power eliminating need system modeling thus decreasing computation.
Язык: Английский
Процитировано
37Nature Photonics, Год журнала: 2024, Номер 18(10), С. 1067 - 1075
Опубликована: Июль 31, 2024
Abstract Optical information processing and computing can potentially offer enhanced performance, scalability energy efficiency. However, achieving nonlinearity—a critical component of computation—remains challenging in the optical domain. Here we introduce a design that leverages multiple-scattering cavity to passively induce nonlinear random mapping with continuous-wave laser at low power. Each scattering event effectively mixes from different areas spatial light modulator, resulting highly between input data output pattern. We demonstrate our retains vital even when readout dimensionality is reduced, thereby enabling compression. This capability allows platforms efficient solutions across applications. design’s efficacy tasks, including classification, image reconstruction, keypoint detection object detection, all which are achieved through compression combined digital decoder. In particular, high performance extreme ratios observed real-time pedestrian detection. Our findings open pathways for novel algorithms unconventional architectural designs computing.
Язык: Английский
Процитировано
17Nature, Год журнала: 2024, Номер 632(8024), С. 280 - 286
Опубликована: Авг. 7, 2024
Optical computing promises to improve the speed and energy efficiency of machine learning applications
Язык: Английский
Процитировано
17Nature Communications, Год журнала: 2024, Номер 15(1)
Опубликована: Апрель 30, 2024
Abstract Ising machines, which are hardware implementations of the model coupled spins, have been influential in development unsupervised learning algorithms at origins Artificial Intelligence (AI). However, their application to AI has limited due complexities matching supervised training methods with machine physics, even though these essential for achieving high accuracy. In this study, we demonstrate an efficient approach train machines a way through Equilibrium Propagation algorithm, comparable results software-based implementations. We employ quantum annealing procedure D-Wave fully-connected neural network on MNIST dataset. Furthermore, that machine’s connectivity supports convolution operations, enabling compact convolutional minimal spins per neuron. Our findings establish as promising trainable platform AI, potential enhance applications.
Язык: Английский
Процитировано
15IEEE Journal on Selected Areas in Communications, Год журнала: 2024, Номер 42(8), С. 2026 - 2038
Опубликована: Май 9, 2024
Wireless networks-on-chip (WNoCs) are an enticing complementary interconnect technology for multi-core chips but face severe resource constraints. Being limited to simple on-off-keying modulation, the reverberant nature of chip enclosure imposes limits on allowed modulation speeds in sight inter-symbol interference, casting doubts competitiveness WNoCs as technology. Fortunately, this vexing problem was recently overcome by parametrizing on-chip radio environment with a reconfigurable intelligent surface (RIS). By suitably configuring RIS, selected channel impulse responses (CIRs) can be tuned (almost) pulse-like despite rich scattering thanks judiciously tailored multi-bounce path interferences. However, exploration "over-the-air" (OTA) equalization is thwarted (i) overwhelming complexity propagation environment, and (ii) non-linear dependence CIR RIS configuration, requiring costly lengthy full-wave simulation every optimization step. Here, we show that reduced-basis physics-compliant model RIS-parametrized calibrated single simulation. Thereby, unlock possibility predicting any configuration almost instantaneously without additional We leverage new tool systematically explore OTA regarding optimal choice delay time RIS-shaped CIR's peak. also study simultaneous multiple wireless links broadcasting conduct performance evaluation terms bit error rate. Looking forward, introduced tools will enable efficient various types analog computing WNoCs.
Язык: Английский
Процитировано
11Physical Review Letters, Год журнала: 2025, Номер 134(1)
Опубликована: Янв. 7, 2025
The rapidly increasing computational demands for artificial intelligence (AI) have spurred the exploration of computing principles beyond conventional digital computers. Physical neural networks (PNNs) offer efficient neuromorphic information processing by harnessing innate power physical processes; however, training their weight parameters is computationally expensive. We propose a approach substantially reducing this cost. Our merges an optimal control method continuous-time dynamical systems with biologically plausible method-direct feedback alignment. In addition to reduction time, achieves robust even under measurement errors and noise without requiring detailed system information. effectiveness was numerically experimentally verified in optoelectronic delay system. significantly extends range practically usable as PNNs.
Язык: Английский
Процитировано
1Light Science & Applications, Год журнала: 2024, Номер 13(1)
Опубликована: Июль 23, 2024
Abstract Nonlinear encoding of optical information can be achieved using various forms data representation. Here, we analyze the performances different nonlinear strategies that employed in diffractive processors based on linear materials and shed light their utility performance gaps compared to state-of-the-art digital deep neural networks. For a comprehensive evaluation, used datasets compare statistical inference simpler-to-implement involve, e.g., phase encoding, against repetition-based strategies. We show repetition within volume (e.g., through an cavity or cascaded introduction input data) causes loss universal transformation capability processor. Therefore, blocks cannot provide analogs fully connected convolutional layers commonly However, they still effectively trained for specific tasks achieve enhanced accuracy, benefiting from information. Our results also reveal without provides simpler strategy with comparable accuracy processors. analyses conclusions would broad interest explore push-pull relationship between material-based systems visual
Язык: Английский
Процитировано
8